For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo struct

KVLoad2CvtConsumer

struct KVLoad2CvtConsumer[dtype: DType, config: MLA_SM100_Decode_Config]

Consumer side of the FP8-to-BF16 load-and-convert KV pipeline.

Parameters

dtype (DType): Element type of the FP8 KV tiles stored in SMEM.
config (MLA_SM100_Decode_Config): Decode config supplying num_kv_stages, BN_QK, and q_depth for pipeline stage sizing.

Fields

pipe (KVLoad2CvtConsumer[dtype, config].PipeT):
smem_upper_fp8 (Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED, _safe=False]):

Implemented traits

AnyType, Copyable, ImplicitlyCopyable, ImplicitlyDeletable, Movable, RegisterPassable, TrivialRegisterPassable

`comptime` members

`bf16_stage_elems`

comptime bf16_stage_elems = (config * config)

`fp8_stage_stride_elems`

comptime fp8_stage_stride_elems = (Int(2) * Int((mul config.BN_QK, config.q_depth)))

`PipeT`

comptime PipeT = KVPipelineGeneric[config.num_kv_stages, Int(1), (Int(1) + Int(32) if (xor (eq (xor (lt config.scale_block_size, 1), True), False), True) else Int(0)), (_resolve_warpgroup_size() + Int(2))]

Methods

`init`

def __init__(pipe: KVPipelineGeneric[config.num_kv_stages, Int(1), (Int(1) + Int(32) if (xor (eq (xor (lt config.scale_block_size, 1), True), False), True) else Int(0)), (_resolve_warpgroup_size() + Int(2))], smem_upper_fp8: Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED, _safe=False]) -> Self

`stage_base_ptr`

def stage_base_ptr(self) -> Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED, _safe=False]

Returns:

Pointer[Scalar[dtype], MutAnyOrigin, address_space=AddressSpace.SHARED, _safe=False]

`wait`

def wait(self)

`release_all`

def release_all(mut self)

Parameters​

Fields​

Implemented traits​

comptime members​

bf16_stage_elems​

fp8_stage_stride_elems​

PipeT​

Methods​

__init__​

stage_base_ptr​

wait​

release_all​